Comparative Analysis of CNN Architectures for Object Recognition
L Ananthakrishanan, Pratik Kumar, M. Harshith, Anshul, Muralitharan Krishanan, Aparna Mohanty · 2025
The primary objective of this research is to conduct a comparative analysis of three convolution neural network (CNN) architectures-LeNet, Simple Custom CNN, and ResNet with three standard image classification datasets-MNIST, CIFAR-10, CIFAR-100. Different level of complexity and diversity are fed into each dataset, offering a comprehensive evaluation of the models' effectiveness in various object recognition scenarios. This research aims to examine how architectural complexity, depth, and design influence model performance on datasets of escalating difficulty. Through the comparison of a traditional architecture (LeNet), optimized custom CNN, and a deep residual model (ResNet), this research focused to highlight the trade-offs between simplicity, accuracy, and generalization. The primary objective is to identify which model excels for particular tasks and under which circumstances, providing insights that could guide mode selection and development in practical applications. More on, this research seeks to highlight the significance of dataset traits in model efficacy and illustrate that increased model complexity does not necessarily lead to improved outcomes, particularly with simpler datasets such as MNIST.